feat(speculative): support Gemma-4-E4B as a streaming DFlash/DSpark target - #2186

Open

h-guo18 wants to merge 9 commits into

mainfrom

haoguo/gemma4-e4b-support

Commits on Aug 13, 2026

feat(speculative): support Gemma-4-E4B as a streaming DFlash/DSpark target
h-guo18
committed
1d321b2

Commits on Aug 18, 2026

feat(speculative): Gemma4-style DSpark draft block (sandwich norms, k_eq_v)
h-guo18
committed
7b56dd6
feat(speculative): Gemma4-shaped DFlash/DSpark draft layer
h-guo18
committed
1dfab24
fix(speculative): export Gemma4 per-layer attention fields; harden the converter
h-guo18
committed
b3dd611
docs(speculative): pin down Gemma4 DSpark draft attention semantics
h-guo18
committed
e9448bb
feat(speculative): Gemma4 DSpark sliding-window (SWA) draft variant
h-guo18
committed
5ccbef4
example(launcher): Gemma-4-E4B DSpark SWA streaming smoke
h-guo18
committed
b27b885

Commits on Aug 19, 2026

example(launcher): Gemma-4-E4B DSpark full 2-epoch streaming run
h-guo18
committed
e82d1ca

Commits on Aug 24, 2026

fix(speculative): preserve the base embedding scale in FakeBaseModel
h-guo18
committed
b7bd27d